Merge pull request #2730 from bakhtiar-id/sync-neuralwatt-260622

Update Neuralwatt models: remove dead GLM 5.1 entries, add GLM 5.2 variants (fast, short, short & fast), add cache-read pricing, and add 5 flex tiers

providers/neuralwatt/README.md

@@ -12,22 +12,33 @@ Model Categories
 
 Reasoning Models (with interleaved thinking):
 - glm-5.2 — GLM 5.2, reasoning enabled
-- zai-org/GLM-5.1-FP8 — GLM 5.1 FP8, reasoning enabled
+- glm-5.2-short — GLM 5.2 Short, reasoning enabled
 - moonshotai/Kimi-K2.5 — Kimi K2.5, reasoning + image input
 - moonshotai/Kimi-K2.6 — Kimi K2.6, reasoning + image input
+- moonshotai/Kimi-K2.7-Code — Kimi K2.7 Code, reasoning + image input
 - Qwen/Qwen3.5-397B-A17B-FP8 — Qwen3.5 397B, reasoning enabled
 - Qwen/Qwen3.6-35B-A3B — Qwen3.6 35B A3B, reasoning enabled
 
 Fast Variants (optimized for speed, non-reasoning):
-- glm-5-fast — GLM 5 Fast
-- glm-5.1-fast — GLM 5.1 Fast
+- glm-5.2-fast — GLM 5.2 Fast
+- glm-5.2-short-fast — GLM 5.2 Short Fast
 - kimi-k2.5-fast — Kimi K2.5 Fast, image input
 - kimi-k2.6-fast — Kimi K2.6 Fast, image input
 - qwen3.5-397b-fast — Qwen3.5 397B Fast
 - qwen3.6-35b-fast — Qwen3.6 35B Fast
 
+Flex Variants (streaming required, discounted):
+- glm-5.2-flex — GLM 5.2 Flex, reasoning enabled
+- glm-5.2-short-flex — GLM 5.2 Short Flex, reasoning enabled
+- glm-5.2-short-fast-flex — GLM 5.2 Short Fast Flex
+- kimi-k2.6-flex — Kimi K2.6 Flex, reasoning + image input
+- kimi-k2.7-code-flex — Kimi K2.7 Code Flex, reasoning + image input
+
 Notes
-- Model IDs, pricing, and limits sourced directly from the Neuralwatt API
+- Standard model IDs, pricing, and limits are sourced directly from the Neuralwatt API; flex pricing applies the official 0.5 docs multiplier to the corresponding standard rates
+- Cache reads are billed at 25% of the input token price; there is no separate cache-write charge
 - Neuralwatt provides real-time energy consumption data (Joules/kWh) per request
 - "Fast" variants are optimized for lower latency without reasoning
+- Flex requires streaming; non-streaming requests fall back to the standard tier
+- Official Neuralwatt docs currently describe Flex as a 50% token-pricing discount, including cached-input billing
 - Vision models support image input via OpenAI-compatible API

providers/neuralwatt/models/Qwen/Qwen3.5-397B-A17B-FP8.toml

@@ -12,6 +12,7 @@ open_weights = true
 [cost]
 input = 0.69
 output = 4.14
+cache_read = 0.1725
 
 [limit]
 context = 262_128
@@ -22,4 +23,4 @@ input = ["text"]
 output = ["text"]
 
 [interleaved]
-field = "reasoning_content"
+field = "reasoning_content"

providers/neuralwatt/models/Qwen/Qwen3.6-35B-A3B.toml

@@ -12,6 +12,7 @@ open_weights = true
 [cost]
 input = 0.29
 output = 1.15
+cache_read = 0.0725
 
 [limit]
 context = 131_056
@@ -22,4 +23,4 @@ input = ["text", "image"]
 output = ["text"]
 
 [interleaved]
-field = "reasoning_content"
+field = "reasoning_content"

providers/neuralwatt/models/glm-5.1-fast.toml

@@ -1,21 +0,0 @@
-name = "GLM 5.1 Fast"
-family = "glm"
-release_date = "2026-04-07"
-last_updated = "2026-04-07"
-attachment = false
-reasoning = false
-temperature = true
-tool_call = true
-open_weights = true
-
-[cost]
-input = 1.1
-output = 3.6
-
-[limit]
-context = 202_736
-output = 202_736
-
-[modalities]
-input = ["text"]
-output = ["text"]

providers/neuralwatt/models/glm-5.2-fast.toml

@@ -0,0 +1,22 @@
+name = "GLM 5.2 Fast"
+family = "glm"
+release_date = "2026-06-17"
+last_updated = "2026-06-17"
+attachment = false
+reasoning = false
+temperature = true
+tool_call = true
+open_weights = true
+
+[cost]
+input = 1.45
+output = 4.5
+cache_read = 0.3625
+
+[limit]
+context = 1_048_560
+output = 1_048_560
+
+[modalities]
+input = ["text"]
+output = ["text"]

providers/neuralwatt/models/glm-5.2-flex.toml

@@ -0,0 +1,27 @@
+name = "GLM 5.2 Flex"
+family = "glm"
+release_date = "2026-06-17"
+last_updated = "2026-06-17"
+attachment = false
+reasoning = true
+reasoning_options = [
+  { type = "toggle" },
+  { type = "effort", values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"] },
+  { type = "budget_tokens" },
+]
+temperature = true
+tool_call = true
+open_weights = true
+
+[cost]
+input = 0.725
+output = 2.25
+cache_read = 0.18125
+
+[limit]
+context = 1_048_560
+output = 1_048_560
+
+[modalities]
+input = ["text"]
+output = ["text"]

providers/neuralwatt/models/glm-5.2-short-fast-flex.toml

@@ -1,7 +1,7 @@
-name = "GLM 5 Fast"
+name = "GLM 5.2 Short Fast Flex"
 family = "glm"
-release_date = "2026-04-07"
-last_updated = "2026-04-07"
+release_date = "2026-06-17"
+last_updated = "2026-06-17"
 attachment = false
 reasoning = false
 temperature = true
@@ -9,12 +9,13 @@ tool_call = true
 open_weights = true
 
 [cost]
-input = 1.1
-output = 3.6
+input = 0.725
+output = 2.25
+cache_read = 0.18125
 
 [limit]
-context = 202_736
-output = 202_736
+context = 199_984
+output = 199_984
 
 [modalities]
 input = ["text"]

providers/neuralwatt/models/glm-5.2-short-fast.toml

@@ -0,0 +1,22 @@
+name = "GLM 5.2 Short Fast"
+family = "glm"
+release_date = "2026-06-17"
+last_updated = "2026-06-17"
+attachment = false
+reasoning = false
+temperature = true
+tool_call = true
+open_weights = true
+
+[cost]
+input = 1.45
+output = 4.5
+cache_read = 0.3625
+
+[limit]
+context = 199_984
+output = 199_984
+
+[modalities]
+input = ["text"]
+output = ["text"]

providers/neuralwatt/models/glm-5.2-short-flex.toml

@@ -0,0 +1,27 @@
+name = "GLM 5.2 Short Flex"
+family = "glm"
+release_date = "2026-06-17"
+last_updated = "2026-06-17"
+attachment = false
+reasoning = true
+reasoning_options = [
+  { type = "toggle" },
+  { type = "effort", values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"] },
+  { type = "budget_tokens" },
+]
+temperature = true
+tool_call = true
+open_weights = true
+
+[cost]
+input = 0.725
+output = 2.25
+cache_read = 0.18125
+
+[limit]
+context = 199_984
+output = 199_984
+
+[modalities]
+input = ["text"]
+output = ["text"]

providers/neuralwatt/models/glm-5.2-short.toml

@@ -1,10 +1,14 @@
-name = "GLM 5.2 short"
+name = "GLM 5.2 Short"
 family = "glm"
 release_date = "2026-06-17"
 last_updated = "2026-06-17"
 attachment = false
 reasoning = true
-reasoning_options = [{ type = "effort", values = ["minimal", "low", "medium", "high", "xhigh"] }]
+reasoning_options = [
+  { type = "toggle" },
+  { type = "effort", values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"] },
+  { type = "budget_tokens" },
+]
 temperature = true
 tool_call = true
 open_weights = true
@@ -12,10 +16,11 @@ open_weights = true
 [cost]
 input = 1.45
 output = 4.5
+cache_read = 0.3625
 
 [limit]
-context = 200_000
-output = 200_000
+context = 199_984
+output = 199_984
 
 [modalities]
 input = ["text"]

providers/neuralwatt/models/glm-5.2.toml

@@ -4,7 +4,11 @@ release_date = "2026-06-17"
 last_updated = "2026-06-17"
 attachment = false
 reasoning = true
-reasoning_options = [{ type = "effort", values = ["minimal", "low", "medium", "high", "xhigh"] }]
+reasoning_options = [
+  { type = "toggle" },
+  { type = "effort", values = ["none", "minimal", "low", "medium", "high", "xhigh", "max"] },
+  { type = "budget_tokens" },
+]
 temperature = true
 tool_call = true
 open_weights = true
@@ -12,6 +16,7 @@ open_weights = true
 [cost]
 input = 1.45
 output = 4.5
+cache_read = 0.3625
 
 [limit]
 context = 1_048_560

providers/neuralwatt/models/kimi-k2.5-fast.toml

@@ -11,6 +11,7 @@ open_weights = true
 [cost]
 input = 0.52
 output = 2.59
+cache_read = 0.13
 
 [limit]
 context = 262_128
@@ -18,4 +19,4 @@ output = 262_128
 
 [modalities]
 input = ["text", "image"]
-output = ["text"]
+output = ["text"]

providers/neuralwatt/models/kimi-k2.6-fast.toml

@@ -11,6 +11,7 @@ open_weights = true
 [cost]
 input = 0.69
 output = 3.22
+cache_read = 0.1725
 
 [limit]
 context = 262_128
@@ -18,6 +19,4 @@ output = 262_128
 
 [modalities]
 input = ["text", "image"]
-output = ["text"]
-
-
+output = ["text"]

providers/neuralwatt/models/kimi-k2.6-flex.toml

@@ -0,0 +1,26 @@
+name = "Kimi K2.6 Flex"
+family = "kimi-k2"
+release_date = "2026-04-21"
+last_updated = "2026-04-21"
+attachment = true
+reasoning = true
+reasoning_options = [{ type = "toggle" }]
+temperature = true
+tool_call = true
+open_weights = true
+
+[cost]
+input = 0.345
+output = 1.61
+cache_read = 0.08625
+
+[limit]
+context = 262_128
+output = 262_128
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]
+
+[interleaved]
+field = "reasoning_content"

providers/neuralwatt/models/kimi-k2.7-code-flex.toml

@@ -0,0 +1,11 @@
+base_model = "moonshotai/kimi-k2.7-code"
+name = "Kimi K2.7 Code Flex"
+
+[cost]
+input = 0.475
+output = 2
+cache_read = 0.11875
+
+[modalities]
+input = ["text", "image"]
+output = ["text"]

providers/neuralwatt/models/moonshotai/Kimi-K2.5.toml

@@ -12,6 +12,7 @@ open_weights = true
 [cost]
 input = 0.52
 output = 2.59
+cache_read = 0.13
 
 [limit]
 context = 262_128
@@ -22,4 +23,4 @@ input = ["text", "image"]
 output = ["text"]
 
 [interleaved]
-field = "reasoning_content"
+field = "reasoning_content"

providers/neuralwatt/models/moonshotai/Kimi-K2.6.toml

@@ -12,6 +12,7 @@ open_weights = true
 [cost]
 input = 0.69
 output = 3.22
+cache_read = 0.1725
 
 [limit]
 context = 262_128
@@ -22,4 +23,4 @@ input = ["text", "image"]
 output = ["text"]
 
 [interleaved]
-field = "reasoning_content"
+field = "reasoning_content"

providers/neuralwatt/models/moonshotai/Kimi-K2.7-Code.toml

@@ -4,7 +4,8 @@ reasoning_options = []
 [cost]
 input = 0.95
 output = 4
+cache_read = 0.2375
 
 [modalities]
 input = ["text", "image"]
-output = ["text"]
+output = ["text"]

providers/neuralwatt/models/qwen3.5-397b-fast.toml

@@ -11,6 +11,7 @@ open_weights = true
 [cost]
 input = 0.69
 output = 4.14
+cache_read = 0.1725
 
 [limit]
 context = 262_128
@@ -18,4 +19,4 @@ output = 262_128
 
 [modalities]
 input = ["text"]
-output = ["text"]
+output = ["text"]

providers/neuralwatt/models/qwen3.6-35b-fast.toml

@@ -11,6 +11,7 @@ open_weights = true
 [cost]
 input = 0.29
 output = 1.15
+cache_read = 0.0725
 
 [limit]
 context = 131_056
@@ -18,4 +19,4 @@ output = 131_056
 
 [modalities]
 input = ["text", "image"]
-output = ["text"]
+output = ["text"]

providers/neuralwatt/models/zai-org/GLM-5.1-FP8.toml

@@ -1,25 +0,0 @@
-name = "GLM 5.1 FP8"
-family = "glm"
-release_date = "2026-04-07"
-last_updated = "2026-04-07"
-attachment = false
-reasoning = true
-reasoning_options = [{ type = "toggle" }]
-temperature = true
-tool_call = true
-open_weights = true
-
-[cost]
-input = 1.1
-output = 3.6
-
-[limit]
-context = 202_736
-output = 202_736
-
-[modalities]
-input = ["text"]
-output = ["text"]
-
-[interleaved]
-field = "reasoning_content"

providers/neuralwatt/provider.toml

@@ -5,6 +5,12 @@ npm = "@ai-sdk/openai-compatible"
 # `thinking_token_budget` (no bounds or disable sentinel documented). Native
 # toggles use `chat_template_kwargs.enable_thinking = true|false`; GLM-5.2 also
 # accepts `reasoning_effort`, with the model-specific normalization documented.
+# Flex tier uses either a `-flex` model ID or `service_tier = "flex"` and
+# requires `stream = true`; non-streaming requests fall through to the standard
+# tier. The official Energy Methodology docs currently state a 0.5 flex pricing
+# multiplier.
 # https://portal.neuralwatt.com/docs/api/chat-completions (accessed 2026-06-25)
+# https://portal.neuralwatt.com/docs/guides/flex-tier
+# https://portal.neuralwatt.com/docs/energy-methodology (accessed 2026-06-28)
 api = "https://api.neuralwatt.com/v1"
 doc = "https://portal.neuralwatt.com/docs"