|
10 | 10 | # Capability matrix: |
11 | 11 | # Model | Reasoning | Tools | Notes |
12 | 12 | # -------------------------|-----------|-------|--------------------------- |
13 | | -# gpt-oss-120b-32k | ✓ | ✓ | OpenAI MoE (5.1B active), ~56GB, fits 2x5090 |
14 | | -# gpt-oss-20b-32k | ✓ | ✓ | OpenAI MoE, 14GB, reasoning + tools |
15 | | -# qwen3-coder:30b | ✓ | ✓ | MoE (3.3B active), both work! |
| 13 | +# gpt-oss-* | ✓ | ✓ | DISABLED - OpenCode ZodError (issue #7439) |
| 14 | +# qwen3-coder-30b-32k | ✓ | ✓ | MoE (3.3B active), both work! |
16 | 15 | # qwen3:32b | ✓ | ~ | Tools buggy in Ollama |
17 | 16 | # llama3.3:70b | ✗ | ✓ | Reliable tools, no thinking |
18 | 17 | # llama3.1-abliterated:70b | ✗ | ✓ | Uncensored, reliable tools |
|
35 | 34 | "$schema" = "https://opencode.ai/config.json"; |
36 | 35 | provider = { |
37 | 36 | ollama = { |
38 | | - # Use ollama-ai-provider-v2 for native reasoning/thinking support |
39 | | - # @ai-sdk/openai-compatible can't parse the non-standard 'reasoning' field |
40 | | - npm = "ollama-ai-provider-v2"; |
| 37 | + # NOTE: gpt-oss models have streaming response format issues with OpenCode |
| 38 | + # The finishReason is returned as object instead of string |
| 39 | + # See: https://github.com/anomalyco/opencode/issues/7439 |
| 40 | + npm = "@ai-sdk/openai-compatible"; |
41 | 41 | name = "Ollama (local)"; |
42 | 42 | options = { |
43 | | - baseURL = "http://localhost:11434/api"; |
| 43 | + baseURL = "http://localhost:11434/v1"; |
44 | 44 | }; |
45 | 45 | models = { |
46 | | - # === GPT-OSS - OpenAI's open-weight MoE models === |
47 | | - # Apache 2.0 license, reasoning + tools, configurable reasoning effort |
| 46 | + # === GPT-OSS - DISABLED: OpenCode streaming compatibility issue === |
| 47 | + # OpenAI's open-weight MoE models (Apache 2.0, reasoning + tools) |
48 | 48 | # Docs: https://ollama.com/library/gpt-oss |
49 | | - |
50 | | - # GPT-OSS 120B - OpenAI's flagship open model |
51 | | - # MoE: 117B total params, 5.1B active, MXFP4 quantized (~56GB weights) |
52 | | - # On 2x 5090 (64GB): fits cleanly, 32k context saturates VRAM |
53 | | - # IMPORTANT: Create variant first: |
54 | | - # ollama run gpt-oss:120b |
55 | | - # /set parameter num_ctx 32768 |
56 | | - # /save gpt-oss-120b-32k |
57 | | - # /bye |
58 | | - "gpt-oss-120b-32k" = { |
59 | | - name = "GPT-OSS 120B 32k (local)"; |
60 | | - reasoning = true; |
61 | | - tool_call = true; |
62 | | - options = { |
63 | | - extraBody = { |
64 | | - think = "high"; # enable reasoning mode |
65 | | - }; |
66 | | - }; |
67 | | - limit = { |
68 | | - context = 32768; # max for 64GB VRAM (model + KV cache) |
69 | | - output = 8192; |
70 | | - }; |
71 | | - }; |
72 | | - # GPT-OSS 20B - smaller variant, fits easily on 64GB |
73 | | - # MoE architecture, 14GB weights, excellent headroom for large context |
74 | | - # IMPORTANT: Create variant first: |
75 | | - # ollama run gpt-oss:20b |
76 | | - # /set parameter num_ctx 32768 |
77 | | - # /save gpt-oss-20b-32k |
78 | | - # /bye |
79 | | - "gpt-oss-20b-32k" = { |
80 | | - name = "GPT-OSS 20B 32k (local)"; |
81 | | - reasoning = true; |
82 | | - tool_call = true; |
83 | | - options = { |
84 | | - extraBody = { |
85 | | - think = "high"; # enable reasoning mode |
86 | | - }; |
87 | | - }; |
88 | | - limit = { |
89 | | - context = 32768; |
90 | | - output = 8192; |
91 | | - }; |
92 | | - }; |
| 49 | + # |
| 50 | + # BUG: gpt-oss returns finishReason as object instead of string in |
| 51 | + # streaming responses, causing ZodError in OpenCode's processor.ts. |
| 52 | + # Affects all providers (@ai-sdk/openai-compatible, ollama-ai-provider-v2). |
| 53 | + # Issue: https://github.com/anomalyco/opencode/issues/7439 |
| 54 | + # |
| 55 | + # Re-enable once OpenCode fixes streaming response parsing. |
| 56 | + # |
| 57 | + # "gpt-oss-120b-32k" = { |
| 58 | + # name = "GPT-OSS 120B 32k (local)"; |
| 59 | + # # MoE: 117B params, 5.1B active, ~56GB MXFP4. Fits 2x5090 w/ 32k ctx. |
| 60 | + # # Create variant: ollama run gpt-oss:120b → /set parameter num_ctx 32768 → /save gpt-oss-120b-32k |
| 61 | + # reasoning = true; |
| 62 | + # tool_call = true; |
| 63 | + # limit = { context = 32768; output = 8192; }; |
| 64 | + # }; |
| 65 | + # "gpt-oss-20b-32k" = { |
| 66 | + # name = "GPT-OSS 20B 32k (local)"; |
| 67 | + # # MoE: 14GB weights, fits easily on 64GB with large context. |
| 68 | + # # Create variant: ollama run gpt-oss:20b → /set parameter num_ctx 32768 → /save gpt-oss-20b-32k |
| 69 | + # reasoning = true; |
| 70 | + # tool_call = true; |
| 71 | + # limit = { context = 32768; output = 8192; }; |
| 72 | + # }; |
93 | 73 |
|
94 | 74 | # === Qwen3-Coder - BOTH reasoning AND reliable tool calling === |
95 | 75 |
|
96 | 76 | # Qwen3-Coder 30B - best for agentic work (reasoning + tools both work) |
97 | 77 | # Unsloth fixed tool calling in Aug 2025 |
98 | | - "qwen3-coder:30b" = { |
99 | | - name = "Qwen3-Coder 30B (local)"; |
| 78 | + # IMPORTANT: Create variant first: |
| 79 | + # ollama run qwen3-coder:30b |
| 80 | + # /set parameter num_ctx 32768 |
| 81 | + # /save qwen3-coder-30b-32k |
| 82 | + # /bye |
| 83 | + "qwen3-coder-30b-32k" = { |
| 84 | + name = "Qwen3-Coder 30B 32k (local)"; |
100 | 85 | reasoning = true; |
101 | 86 | tool_call = true; |
102 | 87 | interleaved = { |
|
0 commit comments